Injected gender cues remain recoverable from name- and pronoun-sanitized LLM résumé assessments

LLM-generated assessments are increasingly stored, shared, and used by downstream systems to make hiring decisions. In order to protect privacy and reduce bias, it is common to obfuscate elements that may reveal personal information about the applicant, such as pronouns. However, it is unknown whether the remaining language of an assessment can still encode this information. We investigate whether the injected gender cue condition (IGC) remains recoverable from obfuscated LLM-generated résumé assessments. Ten models each evaluate four synthetic résumés under counterfactually varied names and pronouns, yielding 8,000 assessments. We find that, as long as the résumé template is known, IGC can be recovered from assessments generated by the same models represented in the classifier's training corpus 69.31% of the time pooled across the four occupations, and up to 76.75% of the time within a single occupation, though recovery does not transfer to an unseen template of the same résumé. The recovery ceiling is 91.93% because generators often emit identical text for candidates of both IGCs.